1 //===- NeonEmitter.cpp - Generate arm_neon.h for use with clang -*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This tablegen backend is responsible for emitting arm_neon.h, which includes
11 // a declaration and definition of each function specified by the ARM NEON
12 // compiler interface. See ARM document DUI0348B.
14 // Each NEON instruction is implemented in terms of 1 or more functions which
15 // are suffixed with the element type of the input vectors. Functions may be
16 // implemented in terms of generic vector operations such as +, *, -, etc. or
17 // by calling a __builtin_-prefixed function which will be handled by clang's
20 // Additional validation code can be generated by this file when runHeader() is
21 // called, rather than the normal run() entry point. A complete set of tests
22 // for Neon intrinsics can be generated by calling the runTests() entry point.
24 //===----------------------------------------------------------------------===//
26 #include "NeonEmitter.h"
27 #include "llvm/ADT/SmallString.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/StringExtras.h"
34 /// ParseTypes - break down a string such as "fQf" into a vector of StringRefs,
35 /// which each StringRef representing a single type declared in the string.
36 /// for "fQf" we would end up with 2 StringRefs, "f", and "Qf", representing
37 /// 2xfloat and 4xfloat respectively.
38 static void ParseTypes(Record *r, std::string &s,
39 SmallVectorImpl<StringRef> &TV) {
40 const char *data = s.data();
43 for (unsigned i = 0, e = s.size(); i != e; ++i, ++len) {
44 if (data[len] == 'P' || data[len] == 'Q' || data[len] == 'U')
56 throw TGError(r->getLoc(),
57 "Unexpected letter: " + std::string(data + len, 1));
60 TV.push_back(StringRef(data, len + 1));
66 /// Widen - Convert a type code into the next wider type. char -> short,
67 /// short -> int, etc.
68 static char Widen(const char t) {
78 default: throw "unhandled type in widen!";
83 /// Narrow - Convert a type code into the next smaller type. short -> char,
84 /// float -> half float, etc.
85 static char Narrow(const char t) {
95 default: throw "unhandled type in narrow!";
100 /// For a particular StringRef, return the base type code, and whether it has
101 /// the quad-vector, polynomial, or unsigned modifiers set.
102 static char ClassifyType(StringRef ty, bool &quad, bool &poly, bool &usgn) {
106 if (ty[off] == 'Q') {
112 if (ty[off] == 'P') {
117 // remember unsigned.
118 if (ty[off] == 'U') {
123 // base type to get the type string for.
127 /// ModType - Transform a type code and its modifiers based on a mod code. The
128 /// mod code definitions may be found at the top of arm_neon.td.
129 static char ModType(const char mod, char type, bool &quad, bool &poly,
130 bool &usgn, bool &scal, bool &cnst, bool &pntr) {
203 /// TypeString - for a modifier and type, generate the name of the typedef for
204 /// that type. QUc -> uint8x8_t.
205 static std::string TypeString(const char mod, StringRef typestr) {
218 // base type to get the type string for.
219 char type = ClassifyType(typestr, quad, poly, usgn);
221 // Based on the modifying character, change the type and width if necessary.
222 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
231 s += poly ? "poly8" : "int8";
234 s += quad ? "x16" : "x8";
237 s += poly ? "poly16" : "int16";
240 s += quad ? "x8" : "x4";
246 s += quad ? "x4" : "x2";
252 s += quad ? "x2" : "x1";
258 s += quad ? "x8" : "x4";
264 s += quad ? "x4" : "x2";
267 throw "unhandled type!";
278 // Append _t, finishing the type string typedef type.
290 /// BuiltinTypeString - for a modifier and type, generate the clang
291 /// BuiltinsARM.def prototype code for the function. See the top of clang's
292 /// Builtins.def for a description of the type strings.
293 static std::string BuiltinTypeString(const char mod, StringRef typestr,
294 ClassKind ck, bool ret) {
307 // base type to get the type string for.
308 char type = ClassifyType(typestr, quad, poly, usgn);
310 // Based on the modifying character, change the type and width if necessary.
311 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
313 // All pointers are void* pointers. Change type to 'v' now.
319 // Treat half-float ('h') types as unsigned short ('s') types.
324 usgn = usgn | poly | ((ck == ClassI || ck == ClassW) && scal && type != 'f');
331 else if (type == 'c')
332 s.push_back('S'); // make chars explicitly signed
334 if (type == 'l') // 64-bit long
346 // Since the return value must be one type, return a vector type of the
347 // appropriate width which we will bitcast. An exception is made for
348 // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
349 // fashion, storing them to a pointer arg.
351 if (mod >= '2' && mod <= '4')
352 return "vv*"; // void result with void* first argument
353 if (mod == 'f' || (ck != ClassB && type == 'f'))
354 return quad ? "V4f" : "V2f";
355 if (ck != ClassB && type == 's')
356 return quad ? "V8s" : "V4s";
357 if (ck != ClassB && type == 'i')
358 return quad ? "V4i" : "V2i";
359 if (ck != ClassB && type == 'l')
360 return quad ? "V2LLi" : "V1LLi";
362 return quad ? "V16Sc" : "V8Sc";
365 // Non-return array types are passed as individual vectors.
367 return quad ? "V16ScV16Sc" : "V8ScV8Sc";
369 return quad ? "V16ScV16ScV16Sc" : "V8ScV8ScV8Sc";
371 return quad ? "V16ScV16ScV16ScV16Sc" : "V8ScV8ScV8ScV8Sc";
373 if (mod == 'f' || (ck != ClassB && type == 'f'))
374 return quad ? "V4f" : "V2f";
375 if (ck != ClassB && type == 's')
376 return quad ? "V8s" : "V4s";
377 if (ck != ClassB && type == 'i')
378 return quad ? "V4i" : "V2i";
379 if (ck != ClassB && type == 'l')
380 return quad ? "V2LLi" : "V1LLi";
382 return quad ? "V16Sc" : "V8Sc";
385 /// MangleName - Append a type or width suffix to a base neon function name,
386 /// and insert a 'q' in the appropriate location if the operation works on
387 /// 128b rather than 64b. E.g. turn "vst2_lane" into "vst2q_lane_f32", etc.
388 static std::string MangleName(const std::string &name, StringRef typestr,
390 if (name == "vcvt_f32_f16")
396 char type = ClassifyType(typestr, quad, poly, usgn);
398 std::string s = name;
403 case ClassS: s += poly ? "_p8" : usgn ? "_u8" : "_s8"; break;
404 case ClassI: s += "_i8"; break;
405 case ClassW: s += "_8"; break;
411 case ClassS: s += poly ? "_p16" : usgn ? "_u16" : "_s16"; break;
412 case ClassI: s += "_i16"; break;
413 case ClassW: s += "_16"; break;
419 case ClassS: s += usgn ? "_u32" : "_s32"; break;
420 case ClassI: s += "_i32"; break;
421 case ClassW: s += "_32"; break;
427 case ClassS: s += usgn ? "_u64" : "_s64"; break;
428 case ClassI: s += "_i64"; break;
429 case ClassW: s += "_64"; break;
436 case ClassI: s += "_f16"; break;
437 case ClassW: s += "_16"; break;
444 case ClassI: s += "_f32"; break;
445 case ClassW: s += "_32"; break;
450 throw "unhandled type!";
456 // Insert a 'q' before the first '_' character so that it ends up before
457 // _lane or _n on vector-scalar operations.
459 size_t pos = s.find('_');
460 s = s.insert(pos, "q");
465 // Generate the string "(argtype a, argtype b, ...)"
466 static std::string GenArgs(const std::string &proto, StringRef typestr) {
467 bool define = proto.find('i') != std::string::npos;
473 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
475 // Immediate macro arguments are used directly instead of being assigned
476 // to local temporaries; prepend an underscore prefix to make their
477 // names consistent with the local temporaries.
481 s += TypeString(proto[i], typestr) + " __";
492 // Macro arguments are not type-checked like inline function arguments, so
493 // assign them to local temporaries to get the right type checking.
494 static std::string GenMacroLocals(const std::string &proto, StringRef typestr) {
498 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
499 // Do not create a temporary for an immediate argument.
500 // That would defeat the whole point of using a macro!
501 if (proto[i] == 'i') continue;
503 s += TypeString(proto[i], typestr) + " __";
514 // Use the vmovl builtin to sign-extend or zero-extend a vector.
515 static std::string Extend(StringRef typestr, const std::string &a) {
517 s = MangleName("vmovl", typestr, ClassS);
522 static std::string Duplicate(unsigned nElts, StringRef typestr,
523 const std::string &a) {
526 s = "(" + TypeString('d', typestr) + "){ ";
527 for (unsigned i = 0; i != nElts; ++i) {
537 static std::string SplatLane(unsigned nElts, const std::string &vec,
538 const std::string &lane) {
539 std::string s = "__builtin_shufflevector(" + vec + ", " + vec;
540 for (unsigned i = 0; i < nElts; ++i)
546 static unsigned GetNumElements(StringRef typestr, bool &quad) {
549 char type = ClassifyType(typestr, quad, dummy, dummy);
552 case 'c': nElts = 8; break;
553 case 's': nElts = 4; break;
554 case 'i': nElts = 2; break;
555 case 'l': nElts = 1; break;
556 case 'h': nElts = 4; break;
557 case 'f': nElts = 2; break;
559 throw "unhandled type!";
562 if (quad) nElts <<= 1;
566 // Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
567 static std::string GenOpString(OpKind op, const std::string &proto,
570 unsigned nElts = GetNumElements(typestr, quad);
572 // If this builtin takes an immediate argument, we need to #define it rather
573 // than use a standard declaration, so that SemaChecking can range check
574 // the immediate passed by the user.
575 bool define = proto.find('i') != std::string::npos;
577 std::string ts = TypeString(proto[0], typestr);
588 s += Extend(typestr, "__a") + " + " + Extend(typestr, "__b") + ";";
591 s += "__a + " + Extend(typestr, "__b") + ";";
597 s += Extend(typestr, "__a") + " - " + Extend(typestr, "__b") + ";";
600 s += "__a - " + Extend(typestr, "__b") + ";";
603 s += "__a * " + Duplicate(nElts, typestr, "__b") + ";";
606 s += "__a * " + SplatLane(nElts, "__b", "__c") + ";";
612 s += MangleName("vmull", typestr, ClassS) + "(__a, " +
613 SplatLane(nElts, "__b", "__c") + ");";
616 s += "__a + (__b * " + Duplicate(nElts, typestr, "__c") + ");";
619 s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
622 s += "__a + (__b * __c);";
625 s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, " +
626 Duplicate(nElts, typestr, "__c") + ");";
629 s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, " +
630 SplatLane(nElts, "__c", "__d") + ");";
633 s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, __c);";
636 s += "__a - (__b * " + Duplicate(nElts, typestr, "__c") + ");";
639 s += "__a - (__b * " + SplatLane(nElts, "__c", "__d") + ");";
642 s += "__a - (__b * __c);";
645 s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, " +
646 Duplicate(nElts, typestr, "__c") + ");";
649 s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, " +
650 SplatLane(nElts, "__c", "__d") + ");";
653 s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, __c);";
656 s += MangleName("vqdmull", typestr, ClassS) + "(__a, " +
657 SplatLane(nElts, "__b", "__c") + ");";
660 s += MangleName("vqdmlal", typestr, ClassS) + "(__a, __b, " +
661 SplatLane(nElts, "__c", "__d") + ");";
664 s += MangleName("vqdmlsl", typestr, ClassS) + "(__a, __b, " +
665 SplatLane(nElts, "__c", "__d") + ");";
668 s += MangleName("vqdmulh", typestr, ClassS) + "(__a, " +
669 SplatLane(nElts, "__b", "__c") + ");";
672 s += MangleName("vqrdmulh", typestr, ClassS) + "(__a, " +
673 SplatLane(nElts, "__b", "__c") + ");";
676 s += "(" + ts + ")(__a == __b);";
679 s += "(" + ts + ")(__a >= __b);";
682 s += "(" + ts + ")(__a <= __b);";
685 s += "(" + ts + ")(__a > __b);";
688 s += "(" + ts + ")(__a < __b);";
712 s += "(" + ts + ")__a;";
715 s += "(" + ts + ")__builtin_shufflevector((int64x1_t)__a";
716 s += ", (int64x1_t)__b, 0, 1);";
720 ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 1);";
724 ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 0);";
727 s += Duplicate(nElts, typestr, "__a") + ";";
730 s += SplatLane(nElts, "__a", "__b") + ";";
733 // ((0 & 1) | (~0 & 2))
735 ts = TypeString(proto[1], typestr);
736 s += "((__a & (" + ts + ")__b) | ";
737 s += "(~__a & (" + ts + ")__c));";
740 s += "__builtin_shufflevector(__a, __a";
741 for (unsigned i = 2; i <= nElts; i += 2)
742 for (unsigned j = 0; j != 2; ++j)
743 s += ", " + utostr(i - j - 1);
747 unsigned WordElts = nElts >> (1 + (int)quad);
748 s += "__builtin_shufflevector(__a, __a";
749 for (unsigned i = WordElts; i <= nElts; i += WordElts)
750 for (unsigned j = 0; j != WordElts; ++j)
751 s += ", " + utostr(i - j - 1);
756 unsigned DblWordElts = nElts >> (int)quad;
757 s += "__builtin_shufflevector(__a, __a";
758 for (unsigned i = DblWordElts; i <= nElts; i += DblWordElts)
759 for (unsigned j = 0; j != DblWordElts; ++j)
760 s += ", " + utostr(i - j - 1);
765 std::string abd = MangleName("vabd", typestr, ClassS) + "(__a, __b)";
766 if (typestr[0] != 'U') {
767 // vabd results are always unsigned and must be zero-extended.
768 std::string utype = "U" + typestr.str();
769 s += "(" + TypeString(proto[0], typestr) + ")";
770 abd = "(" + TypeString('d', utype) + ")" + abd;
771 s += Extend(utype, abd) + ";";
773 s += Extend(typestr, abd) + ";";
778 s += "__a + " + MangleName("vabd", typestr, ClassS) + "(__b, __c);";
782 std::string abd = MangleName("vabd", typestr, ClassS) + "(__b, __c)";
783 if (typestr[0] != 'U') {
784 // vabd results are always unsigned and must be zero-extended.
785 std::string utype = "U" + typestr.str();
786 s += "(" + TypeString(proto[0], typestr) + ")";
787 abd = "(" + TypeString('d', utype) + ")" + abd;
788 s += Extend(utype, abd) + ";";
790 s += Extend(typestr, abd) + ";";
795 throw "unknown OpKind!";
801 static unsigned GetNeonEnum(const std::string &proto, StringRef typestr) {
802 unsigned mod = proto[0];
805 if (mod == 'v' || mod == 'f')
815 // Base type to get the type string for.
816 char type = ClassifyType(typestr, quad, poly, usgn);
818 // Based on the modifying character, change the type and width if necessary.
819 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
823 if (quad && proto[1] != 'g')
846 throw "unhandled type!";
852 // Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
853 static std::string GenBuiltin(const std::string &name, const std::string &proto,
854 StringRef typestr, ClassKind ck) {
857 // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
858 // sret-like argument.
859 bool sret = (proto[0] >= '2' && proto[0] <= '4');
861 // If this builtin takes an immediate argument, we need to #define it rather
862 // than use a standard declaration, so that SemaChecking can range check
863 // the immediate passed by the user.
864 bool define = proto.find('i') != std::string::npos;
866 // Check if the prototype has a scalar operand with the type of the vector
867 // elements. If not, bitcasting the args will take care of arg checking.
868 // The actual signedness etc. will be taken care of with special enums.
869 if (proto.find('s') == std::string::npos)
872 if (proto[0] != 'v') {
873 std::string ts = TypeString(proto[0], typestr);
883 s += "return (" + ts + ")";
887 bool splat = proto.find('a') != std::string::npos;
889 s += "__builtin_neon_";
891 // Call the non-splat builtin: chop off the "_n" suffix from the name.
892 std::string vname(name, 0, name.size()-2);
893 s += MangleName(vname, typestr, ck);
895 s += MangleName(name, typestr, ck);
899 // Pass the address of the return variable as the first argument to sret-like
905 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
906 std::string args = std::string(&arg, 1);
908 // Use the local temporaries instead of the macro arguments.
911 bool argQuad = false;
912 bool argPoly = false;
913 bool argUsgn = false;
914 bool argScalar = false;
916 char argType = ClassifyType(typestr, argQuad, argPoly, argUsgn);
917 argType = ModType(proto[i], argType, argQuad, argPoly, argUsgn, argScalar,
920 // Handle multiple-vector values specially, emitting each subvector as an
921 // argument to the __builtin.
922 if (proto[i] >= '2' && proto[i] <= '4') {
923 // Check if an explicit cast is needed.
924 if (argType != 'c' || argPoly || argUsgn)
925 args = (argQuad ? "(int8x16_t)" : "(int8x8_t)") + args;
927 for (unsigned vi = 0, ve = proto[i] - '0'; vi != ve; ++vi) {
928 s += args + ".val[" + utostr(vi) + "]";
938 if (splat && (i + 1) == e)
939 args = Duplicate(GetNumElements(typestr, argQuad), typestr, args);
941 // Check if an explicit cast is needed.
942 if ((splat || !argScalar) &&
943 ((ck == ClassB && argType != 'c') || argPoly || argUsgn)) {
944 std::string argTypeStr = "c";
946 argTypeStr = argType;
948 argTypeStr = "Q" + argTypeStr;
949 args = "(" + TypeString('d', argTypeStr) + ")" + args;
957 // Extra constant integer to hold type class enum for this function, e.g. s8
959 s += ", " + utostr(GetNeonEnum(proto, typestr));
963 if (proto[0] != 'v' && sret) {
972 static std::string GenBuiltinDef(const std::string &name,
973 const std::string &proto,
974 StringRef typestr, ClassKind ck) {
975 std::string s("BUILTIN(__builtin_neon_");
977 // If all types are the same size, bitcasting the args will take care
978 // of arg checking. The actual signedness etc. will be taken care of with
980 if (proto.find('s') == std::string::npos)
983 s += MangleName(name, typestr, ck);
986 for (unsigned i = 0, e = proto.size(); i != e; ++i)
987 s += BuiltinTypeString(proto[i], typestr, ck, i == 0);
989 // Extra constant integer to hold type class enum for this function, e.g. s8
997 static std::string GenIntrinsic(const std::string &name,
998 const std::string &proto,
999 StringRef outTypeStr, StringRef inTypeStr,
1000 OpKind kind, ClassKind classKind) {
1001 assert(!proto.empty() && "");
1002 bool define = proto.find('i') != std::string::npos;
1005 // static always inline + return type
1009 s += "__ai " + TypeString(proto[0], outTypeStr) + " ";
1011 // Function name with type suffix
1012 std::string mangledName = MangleName(name, outTypeStr, ClassS);
1013 if (outTypeStr != inTypeStr) {
1014 // If the input type is different (e.g., for vreinterpret), append a suffix
1015 // for the input type. String off a "Q" (quad) prefix so that MangleName
1016 // does not insert another "q" in the name.
1017 unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
1018 StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
1019 mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
1023 // Function arguments
1024 s += GenArgs(proto, inTypeStr);
1028 s += " __extension__ ({ \\\n ";
1029 s += GenMacroLocals(proto, inTypeStr);
1035 s += GenOpString(kind, proto, outTypeStr);
1037 s += GenBuiltin(name, proto, outTypeStr, classKind);
1046 /// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
1047 /// is comprised of type definitions and function declarations.
1048 void NeonEmitter::run(raw_ostream &OS) {
1050 "/*===---- arm_neon.h - ARM Neon intrinsics ------------------------------"
1053 " * Permission is hereby granted, free of charge, to any person obtaining "
1055 " * of this software and associated documentation files (the \"Software\"),"
1057 " * in the Software without restriction, including without limitation the "
1059 " * to use, copy, modify, merge, publish, distribute, sublicense, "
1061 " * copies of the Software, and to permit persons to whom the Software is\n"
1062 " * furnished to do so, subject to the following conditions:\n"
1064 " * The above copyright notice and this permission notice shall be "
1066 " * all copies or substantial portions of the Software.\n"
1068 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
1070 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
1071 "MERCHANTABILITY,\n"
1072 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
1074 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
1076 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
1078 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
1080 " * THE SOFTWARE.\n"
1082 " *===--------------------------------------------------------------------"
1086 OS << "#ifndef __ARM_NEON_H\n";
1087 OS << "#define __ARM_NEON_H\n\n";
1089 OS << "#ifndef __ARM_NEON__\n";
1090 OS << "#error \"NEON support not enabled\"\n";
1093 OS << "#include <stdint.h>\n\n";
1095 // Emit NEON-specific scalar typedefs.
1096 OS << "typedef float float32_t;\n";
1097 OS << "typedef int8_t poly8_t;\n";
1098 OS << "typedef int16_t poly16_t;\n";
1099 OS << "typedef uint16_t float16_t;\n";
1101 // Emit Neon vector typedefs.
1102 std::string TypedefTypes("cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfPcQPcPsQPs");
1103 SmallVector<StringRef, 24> TDTypeVec;
1104 ParseTypes(0, TypedefTypes, TDTypeVec);
1106 // Emit vector typedefs.
1107 for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1108 bool dummy, quad = false, poly = false;
1109 (void) ClassifyType(TDTypeVec[i], quad, poly, dummy);
1111 OS << "typedef __attribute__((neon_polyvector_type(";
1113 OS << "typedef __attribute__((neon_vector_type(";
1115 unsigned nElts = GetNumElements(TDTypeVec[i], quad);
1116 OS << utostr(nElts) << "))) ";
1120 OS << TypeString('s', TDTypeVec[i]);
1121 OS << " " << TypeString('d', TDTypeVec[i]) << ";\n";
1125 // Emit struct typedefs.
1126 for (unsigned vi = 2; vi != 5; ++vi) {
1127 for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1128 std::string ts = TypeString('d', TDTypeVec[i]);
1129 std::string vs = TypeString('0' + vi, TDTypeVec[i]);
1130 OS << "typedef struct " << vs << " {\n";
1131 OS << " " << ts << " val";
1132 OS << "[" << utostr(vi) << "]";
1134 OS << vs << ";\n\n";
1138 OS << "#define __ai static __attribute__((__always_inline__))\n\n";
1140 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1142 // Emit vmovl, vmull and vabd intrinsics first so they can be used by other
1143 // intrinsics. (Some of the saturating multiply instructions are also
1144 // used to implement the corresponding "_lane" variants, but tablegen
1145 // sorts the records into alphabetical order so that the "_lane" variants
1146 // come after the intrinsics they use.)
1147 emitIntrinsic(OS, Records.getDef("VMOVL"));
1148 emitIntrinsic(OS, Records.getDef("VMULL"));
1149 emitIntrinsic(OS, Records.getDef("VABD"));
1151 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1153 if (R->getName() != "VMOVL" &&
1154 R->getName() != "VMULL" &&
1155 R->getName() != "VABD")
1156 emitIntrinsic(OS, R);
1159 OS << "#undef __ai\n\n";
1160 OS << "#endif /* __ARM_NEON_H */\n";
1163 /// emitIntrinsic - Write out the arm_neon.h header file definitions for the
1164 /// intrinsics specified by record R.
1165 void NeonEmitter::emitIntrinsic(raw_ostream &OS, Record *R) {
1166 std::string name = R->getValueAsString("Name");
1167 std::string Proto = R->getValueAsString("Prototype");
1168 std::string Types = R->getValueAsString("Types");
1170 SmallVector<StringRef, 16> TypeVec;
1171 ParseTypes(R, Types, TypeVec);
1173 OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1175 ClassKind classKind = ClassNone;
1176 if (R->getSuperClasses().size() >= 2)
1177 classKind = ClassMap[R->getSuperClasses()[1]];
1178 if (classKind == ClassNone && kind == OpNone)
1179 throw TGError(R->getLoc(), "Builtin has no class kind");
1181 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1182 if (kind == OpReinterpret) {
1183 bool outQuad = false;
1185 (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1186 for (unsigned srcti = 0, srcte = TypeVec.size();
1187 srcti != srcte; ++srcti) {
1188 bool inQuad = false;
1189 (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1190 if (srcti == ti || inQuad != outQuad)
1192 OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[srcti],
1196 OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[ti],
1203 static unsigned RangeFromType(const char mod, StringRef typestr) {
1204 // base type to get the type string for.
1205 bool quad = false, dummy = false;
1206 char type = ClassifyType(typestr, quad, dummy, dummy);
1207 type = ModType(mod, type, quad, dummy, dummy, dummy, dummy, dummy);
1211 return (8 << (int)quad) - 1;
1214 return (4 << (int)quad) - 1;
1217 return (2 << (int)quad) - 1;
1219 return (1 << (int)quad) - 1;
1221 throw "unhandled type!";
1224 assert(0 && "unreachable");
1228 /// runHeader - Emit a file with sections defining:
1229 /// 1. the NEON section of BuiltinsARM.def.
1230 /// 2. the SemaChecking code for the type overload checking.
1231 /// 3. the SemaChecking code for validation of intrinsic immedate arguments.
1232 void NeonEmitter::runHeader(raw_ostream &OS) {
1233 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1235 StringMap<OpKind> EmittedMap;
1237 // Generate BuiltinsARM.def for NEON
1238 OS << "#ifdef GET_NEON_BUILTINS\n";
1239 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1241 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1245 std::string Proto = R->getValueAsString("Prototype");
1247 // Functions with 'a' (the splat code) in the type prototype should not get
1248 // their own builtin as they use the non-splat variant.
1249 if (Proto.find('a') != std::string::npos)
1252 std::string Types = R->getValueAsString("Types");
1253 SmallVector<StringRef, 16> TypeVec;
1254 ParseTypes(R, Types, TypeVec);
1256 if (R->getSuperClasses().size() < 2)
1257 throw TGError(R->getLoc(), "Builtin has no class kind");
1259 std::string name = R->getValueAsString("Name");
1260 ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1262 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1263 // Generate the BuiltinsARM.def declaration for this builtin, ensuring
1264 // that each unique BUILTIN() macro appears only once in the output
1266 std::string bd = GenBuiltinDef(name, Proto, TypeVec[ti], ck);
1267 if (EmittedMap.count(bd))
1270 EmittedMap[bd] = OpNone;
1276 // Generate the overloaded type checking code for SemaChecking.cpp
1277 OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
1278 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1280 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1284 std::string Proto = R->getValueAsString("Prototype");
1285 std::string Types = R->getValueAsString("Types");
1286 std::string name = R->getValueAsString("Name");
1288 // Functions with 'a' (the splat code) in the type prototype should not get
1289 // their own builtin as they use the non-splat variant.
1290 if (Proto.find('a') != std::string::npos)
1293 // Functions which have a scalar argument cannot be overloaded, no need to
1294 // check them if we are emitting the type checking code.
1295 if (Proto.find('s') != std::string::npos)
1298 SmallVector<StringRef, 16> TypeVec;
1299 ParseTypes(R, Types, TypeVec);
1301 if (R->getSuperClasses().size() < 2)
1302 throw TGError(R->getLoc(), "Builtin has no class kind");
1304 int si = -1, qi = -1;
1305 unsigned mask = 0, qmask = 0;
1306 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1307 // Generate the switch case(s) for this builtin for the type validation.
1308 bool quad = false, poly = false, usgn = false;
1309 (void) ClassifyType(TypeVec[ti], quad, poly, usgn);
1313 qmask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1316 mask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1320 OS << "case ARM::BI__builtin_neon_"
1321 << MangleName(name, TypeVec[si], ClassB)
1322 << ": mask = " << "0x" << utohexstr(mask) << "; break;\n";
1324 OS << "case ARM::BI__builtin_neon_"
1325 << MangleName(name, TypeVec[qi], ClassB)
1326 << ": mask = " << "0x" << utohexstr(qmask) << "; break;\n";
1330 // Generate the intrinsic range checking code for shift/lane immediates.
1331 OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
1332 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1335 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1339 std::string name = R->getValueAsString("Name");
1340 std::string Proto = R->getValueAsString("Prototype");
1341 std::string Types = R->getValueAsString("Types");
1343 // Functions with 'a' (the splat code) in the type prototype should not get
1344 // their own builtin as they use the non-splat variant.
1345 if (Proto.find('a') != std::string::npos)
1348 // Functions which do not have an immediate do not need to have range
1349 // checking code emitted.
1350 size_t immPos = Proto.find('i');
1351 if (immPos == std::string::npos)
1354 SmallVector<StringRef, 16> TypeVec;
1355 ParseTypes(R, Types, TypeVec);
1357 if (R->getSuperClasses().size() < 2)
1358 throw TGError(R->getLoc(), "Builtin has no class kind");
1360 ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1362 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1363 std::string namestr, shiftstr, rangestr;
1365 // Builtins which are overloaded by type will need to have their upper
1366 // bound computed at Sema time based on the type constant.
1367 if (Proto.find('s') == std::string::npos) {
1369 if (R->getValueAsBit("isShift")) {
1370 shiftstr = ", true";
1372 // Right shifts have an 'r' in the name, left shifts do not.
1373 if (name.find('r') != std::string::npos)
1374 rangestr = "l = 1; ";
1376 rangestr += "u = RFT(TV" + shiftstr + ")";
1378 // The immediate generally refers to a lane in the preceding argument.
1379 assert(immPos > 0 && "unexpected immediate operand");
1380 rangestr = "u = " + utostr(RangeFromType(Proto[immPos-1], TypeVec[ti]));
1382 // Make sure cases appear only once by uniquing them in a string map.
1383 namestr = MangleName(name, TypeVec[ti], ck);
1384 if (EmittedMap.count(namestr))
1386 EmittedMap[namestr] = OpNone;
1388 // Calculate the index of the immediate that should be range checked.
1389 unsigned immidx = 0;
1391 // Builtins that return a struct of multiple vectors have an extra
1392 // leading arg for the struct return.
1393 if (Proto[0] >= '2' && Proto[0] <= '4')
1396 // Add one to the index for each argument until we reach the immediate
1397 // to be checked. Structs of vectors are passed as multiple arguments.
1398 for (unsigned ii = 1, ie = Proto.size(); ii != ie; ++ii) {
1399 switch (Proto[ii]) {
1400 default: immidx += 1; break;
1401 case '2': immidx += 2; break;
1402 case '3': immidx += 3; break;
1403 case '4': immidx += 4; break;
1404 case 'i': ie = ii + 1; break;
1407 OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
1408 << ": i = " << immidx << "; " << rangestr << "; break;\n";
1414 /// GenTest - Write out a test for the intrinsic specified by the name and
1415 /// type strings, including the embedded patterns for FileCheck to match.
1416 static std::string GenTest(const std::string &name,
1417 const std::string &proto,
1418 StringRef outTypeStr, StringRef inTypeStr,
1420 assert(!proto.empty() && "");
1423 // Function name with type suffix
1424 std::string mangledName = MangleName(name, outTypeStr, ClassS);
1425 if (outTypeStr != inTypeStr) {
1426 // If the input type is different (e.g., for vreinterpret), append a suffix
1427 // for the input type. String off a "Q" (quad) prefix so that MangleName
1428 // does not insert another "q" in the name.
1429 unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
1430 StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
1431 mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
1434 // Emit the FileCheck patterns.
1435 s += "// CHECK: test_" + mangledName + "\n";
1436 // s += "// CHECK: \n"; // FIXME: + expected instruction opcode.
1438 // Emit the start of the test function.
1439 s += TypeString(proto[0], outTypeStr) + " test_" + mangledName + "(";
1442 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
1443 // Do not create arguments for values that must be immediate constants.
1444 if (proto[i] == 'i')
1446 s += comma + TypeString(proto[i], inTypeStr) + " ";
1452 if (proto[0] != 'v')
1454 s += mangledName + "(";
1456 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
1457 if (proto[i] == 'i') {
1458 // For immediate operands, test the maximum value.
1462 // The immediate generally refers to a lane in the preceding argument.
1463 s += utostr(RangeFromType(proto[i-1], inTypeStr));
1474 /// runTests - Write out a complete set of tests for all of the Neon
1476 void NeonEmitter::runTests(raw_ostream &OS) {
1478 "// RUN: %clang_cc1 -triple thumbv7-apple-darwin \\\n"
1479 "// RUN: -target-cpu cortex-a9 -ffreestanding -S -o - %s | FileCheck %s\n"
1481 "#include <arm_neon.h>\n"
1484 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1485 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1487 std::string name = R->getValueAsString("Name");
1488 std::string Proto = R->getValueAsString("Prototype");
1489 std::string Types = R->getValueAsString("Types");
1490 bool isShift = R->getValueAsBit("isShift");
1492 SmallVector<StringRef, 16> TypeVec;
1493 ParseTypes(R, Types, TypeVec);
1495 OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1496 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1497 if (kind == OpReinterpret) {
1498 bool outQuad = false;
1500 (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1501 for (unsigned srcti = 0, srcte = TypeVec.size();
1502 srcti != srcte; ++srcti) {
1503 bool inQuad = false;
1504 (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1505 if (srcti == ti || inQuad != outQuad)
1507 OS << GenTest(name, Proto, TypeVec[ti], TypeVec[srcti], isShift);
1510 OS << GenTest(name, Proto, TypeVec[ti], TypeVec[ti], isShift);